2.4.2

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -27,6 +27,7 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/argument.h"
 #include "paddle/fluid/string/pretty_log.h"
+#include "paddle/phi/core/errors.h"
 
 namespace paddle {
 namespace inference {
@@ -36,15 +37,6 @@ using string::PrettyLogEndl;
 using string::Style;
 
 IRPassManager::IRPassManager(Argument *argument) {
-  ARGUMENT_CHECK_FIELD(argument, main_program);
-  graph_ = std::unique_ptr<Graph>(new Graph(argument->main_program()));
-  if (argument->Has("scope")) {
-    auto *scope_ptr = argument->scope_ptr();
-    PADDLE_ENFORCE_NOT_NULL(scope_ptr,
-                            platform::errors::PreconditionNotMet(
-                                "The scope ptr should not be nullptr."));
-    graph_->SetNotOwned(framework::ir::kParamScopeAttr, scope_ptr);
-  }
   disable_logs_ = argument->disable_logs();
 
   ARGUMENT_CHECK_FIELD(argument, ir_analysis_passes);
@@ -95,10 +87,14 @@ void IRPassManager::CreatePasses(Argument *argument,
                               argument->tensorrt_tuned_dynamic_shape();
     pass->Set("with_dynamic_shape", new bool(with_dynamic_shape));
 
+    // mixed precision related
     pass->Set("model_precision", new int(argument->model_precision()));
     pass->Set(
         "mixed_black_list",
         new std::unordered_set<std::string>(argument->mixed_black_list()));
+    pass->Set("enable_gpu_mixed", new bool(argument->enable_gpu_mixed()));
+    pass->Set("mixed_precision_mode",
+              new int(argument->mixed_precision_mode()));
 
     if (pass_name == "graph_viz_pass") {
       std::string optim_cache_dir = argument->optim_cache_dir();
@@ -302,42 +298,18 @@ void IRPassManager::CreatePasses(Argument *argument,
 }
 
 std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
-  if (passes_.empty()) {
-    return graph;
-  }
   PADDLE_ENFORCE_NOT_NULL(
-      graph.get(),
-      platform::errors::PreconditionNotMet("Graph cannot be NULL."));
+      graph.get(), platform::errors::InvalidArgument("Graph cannot be null."));
   // Apply all the passes
   for (const auto &pass : passes_) {
     if (pass->Type() != "graph_viz_pass" && !disable_logs_) {
       PrettyLogEndl(Style::H2(), "--- Running IR pass [%s]", pass->Type());
     }
-    // delete_fill_constant_op_pass is not apply under trt dynamic shape
-    if (pass->Type() == "delete_fill_constant_op_pass") {
-      bool use_dynamic = pass->Get<bool>("with_dynamic_shape");
-      if (use_dynamic) continue;
-    }
     graph.reset(pass->Apply(graph.release()));
   }
   return graph;
 }
 
-framework::proto::ProgramDesc IRPassManager::AcquireProgram(
-    std::unique_ptr<Graph> *graph, ProgramDesc *program) const {
-  auto pass =
-      framework::ir::PassRegistry::Instance().Get("graph_to_program_pass");
-
-  // Direct using ProgramDesc desc(argument->main_program()) may cause
-  // incomplete copies of information.
-  ProgramDesc desc;
-  desc.CopyFrom(*program->Proto());
-  pass->SetNotOwned("program", &desc);
-  auto *the_graph = graph->release();
-  graph->reset(pass->Apply(the_graph));
-  return *desc.Proto();
-}
-
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/ir_pass_manager.h

@@ -48,15 +48,9 @@ class IRPassManager final {
 
   std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph);
 
-  framework::proto::ProgramDesc AcquireProgram(std::unique_ptr<Graph> *graph,
-                                               ProgramDesc *program) const;
-
-  framework::ir::Graph &graph() const { return *graph_; }
-
  private:
   void CreatePasses(Argument *argument, const std::vector<std::string> &passes);
 
-  std::unique_ptr<Graph> graph_;
   std::vector<std::unique_ptr<Pass>> passes_;
   bool disable_logs_{false};
 };

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc

@@ -94,14 +94,14 @@ void OutputProcess(framework::ir::Graph *graph,
                   backend,
                   precision,
                   blacklist)) {
-            AddCastOp(graph,
-                      var_node,
-                      next_op,
-                      framework::proto::VarType::FP32,
-                      to_type,
-                      &suffix,
-                      block_desc,
-                      &var_to_cast_op_map);
+            InsertCastOp(graph,
+                         var_node,
+                         next_op,
+                         framework::proto::VarType::FP32,
+                         to_type,
+                         block_desc,
+                         &suffix,
+                         &var_to_cast_op_map);
             var_node->Var()->SetDataType(framework::proto::VarType::FP32);
           }
         }

paddle/fluid/inference/analysis/passes/CMakeLists.txt

@@ -13,7 +13,7 @@ cc_library(
 cc_library(
   convert_to_mixed_precision
   SRCS convert_to_mixed_precision.cc
-  DEPS analysis_pass ir_graph_build_pass)
+  DEPS analysis_pass ir_graph_build_pass auto_mixed_precision_pass)
 cc_library(
   ir_params_sync_among_devices_pass
   SRCS ir_params_sync_among_devices_pass.cc
@@ -30,17 +30,6 @@ cc_library(
   inference_op_replace_pass
   SRCS inference_op_replace_pass.cc
   DEPS analysis_pass graph_to_program_pass)
-if(WITH_TESTING)
-  cc_library(
-    ir_graph_clean_pass
-    SRCS ir_graph_clean_pass.cc
-    DEPS analysis_pass gtest)
-else()
-  cc_library(
-    ir_graph_clean_pass
-    SRCS ir_graph_clean_pass.cc
-    DEPS analysis_pass)
-endif()
 
 cc_library(
   analysis_passes
@@ -52,8 +41,7 @@ cc_library(
        memory_optim_pass
        convert_to_mixed_precision
        inference_op_replace_pass
-       ir_graph_to_program_pass
-       ir_graph_clean_pass)
+       ir_graph_to_program_pass)
 
 set(analysis_deps
     ${analysis_deps} analysis_passes subgraph_detector

paddle/fluid/inference/analysis/passes/convert_to_mixed_precision.cc

@@ -14,807 +14,88 @@
 
 #include "paddle/fluid/inference/analysis/passes/convert_to_mixed_precision.h"
 
-#include <algorithm>
-#include <iterator>
-#include <memory>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <utility>
-
-#include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/executor.h"
-#include "paddle/fluid/framework/framework.pb.h"
-#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/auto_mixed_precision_pass.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
-#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
-#include "paddle/fluid/framework/ir/node.h"
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/framework/var_desc.h"
-#include "paddle/fluid/inference/analysis/argument.h"
-#include "paddle/fluid/inference/analysis/passes/ir_graph_clean_pass.h"
 #include "paddle/fluid/inference/io.h"
-#include "paddle/phi/common/bfloat16.h"
-#include "paddle/phi/common/data_type.h"
-#include "paddle/phi/common/float16.h"
-#include "paddle/phi/common/layout.h"
-#include "paddle/phi/common/place.h"
-#include "paddle/phi/core/tensor_meta.h"
-
-using namespace paddle::framework;  // NOLINT
+#include "paddle/phi/common/backend.h"
 
 namespace paddle {
 namespace inference {
 namespace analysis {
 
-namespace {
-bool PhiKernelSupportPrecision(
-    const std::string& op_type,
+ConvertToMixedPrecisionPass::ConvertToMixedPrecisionPass(
+    const std::string& model_file,
+    const std::string& params_file,
+    const std::string& mixed_model_file,
+    const std::string& mixed_params_file,
+    phi::DataType mixed_precision,
     phi::Backend backend,
-    phi::DataType data_type,
-    phi::DataLayout layout = phi::DataLayout::ALL_LAYOUT) {
-  auto kernels = phi::KernelFactory::Instance().kernels();
-  if (kernels.find(op_type) == kernels.end()) {
-    return false;
-  }
-  phi::KernelKey kernel_key(backend, layout, data_type);
-  return phi::KernelFactory::Instance().HasKernel(op_type, kernel_key);
-}
-
-bool GpuKernelSupportPrecision(
-    const std::string& op_type,
-    phi::DataType data_type,
-    phi::DataLayout layout = phi::DataLayout::ALL_LAYOUT) {
-  auto phi_op_type = phi::TransToPhiKernelName(op_type);
-  bool res = PhiKernelSupportPrecision(
-      phi_op_type, phi::Backend::GPU, data_type, layout);
-  res |= PhiKernelSupportPrecision(
-      phi_op_type, phi::Backend::GPUDNN, data_type, layout);
-
-  if (!res) {
-    auto& all_kernels = OperatorWithKernel::AllOpKernels();
-    auto it = all_kernels.find(op_type);
-    if (it != all_kernels.end()) {
-      for (auto& kern_pair : it->second) {
-        if (platform::is_gpu_place(kern_pair.first.place_) &&
-            kern_pair.first.data_type_ == framework::proto::VarType::FP16) {
-          res = true;
-        }
-      }
-    }
-  }
-  return res;
-}
-
-class ConvertToMixedPrecisionPass {
- public:
-  explicit ConvertToMixedPrecisionPass(
-      const std::string& model_file,
-      const std::string& params_file,
-      const std::string& mixed_model_file,
-      const std::string& mixed_params_file,
-      phi::DataType mixed_precision,
-      phi::Backend backend,
-      bool keep_io_types,
-      std::unordered_set<std::string> black_list)
-      : model_file_(model_file),
-        params_file_(params_file),
-        mixed_model_file_(mixed_model_file),
-        mixed_params_file_(mixed_params_file),
-        mixed_precision_(mixed_precision),
-        backend_(backend),
-        keep_io_types_(keep_io_types),
-        black_list_(black_list),
-        place_(paddle::CPUPlace()),
-        executor_(place_) {
-    black_list_.insert("assign");
-    black_list_.insert("fill_constant");
-    black_list_.insert("assign_value");
-    black_list_.insert("eye");
-    black_list_.insert("fill_any_like");
-    black_list_.insert("fill_constant_batch_size_like");
-  }
-  void Run();
-
- private:
-  void LoadAndPrepare();
-  inline bool NodeVarHasDtype(framework::ir::Node* node);
-  void ConvertAllFp64ToFp32(framework::ir::Graph* graph);
-  void FixCastAttr(framework::ir::Graph* graph);
-  void SaveMixedModel();
-  void ConvertTensorDtype(int block_idx);
-  void ProcessInputNode(bool support_precision,
-                        ir::Node* in_node,
-                        ir::Node* op_node,
-                        int* suffix,
-                        framework::BlockDesc* block_desc,
-                        framework::proto::VarType::Type to_type,
-                        int block_idx);
-
-  void ProcessOutputNode(int block_idx,
-                         ir::Node* var_node,
-                         framework::proto::VarType::Type to_type);
-  inline bool IsFloatVarType(framework::proto::VarType::Type type);
-
-  bool OutShouldNotConvert(ir::Node* var_node);
-  // Just process special cases for weights conversion.
-  bool WeightsShouldNotConvert(ir::Node* var_node);
-
-  // To support multi block, we need to consider a lot of special cases.
-  // Return Node* which first appers in block.
-  framework::ir::Node* GetRealNode(int block_idx, framework::ir::Node* node);
-  void FindVarsInMultiBlock();
-  inline bool VarIsMultiPrecisionOpsOut(int block_idx,
-                                        framework::ir::Node* op_node);
-
- private:
-  // A trick. Patch for strange op, which input name equal to output name, such
-  // as `fused_multi_transformer`
-  void PatchForStrangeOp();
-
- private:
-  std::string model_file_;
-  std::string params_file_;
-  std::string mixed_model_file_;
-  std::string mixed_params_file_;
-  phi::DataType mixed_precision_;
-  phi::Backend backend_;
-  bool keep_io_types_;
-  std::unordered_set<std::string> black_list_;
-  paddle::CPUPlace place_;
-  framework::Executor executor_;
-  framework::Scope scope_;
-
-  std::unordered_map<framework::ir::Node*, framework::ir::Node*> cast_map_;
-  std::unordered_map<std::string,
-                     std::pair<framework::proto::VarType::Type, int>>
-      vars_in_multi_block_map_;
-  std::vector<std::unordered_map<std::string, std::vector<std::string>>>
-      vars_appear_multi_in_one_block_;
-  int suffix_{0};
-
-  std::unique_ptr<framework::ProgramDesc> program_desc_{nullptr};
-  std::unique_ptr<framework::ir::Graph> main_graph_{nullptr};
-  std::vector<framework::ir::Graph*> graphes_;
-};
-
-framework::ir::Node* ConvertToMixedPrecisionPass::GetRealNode(
-    int block_idx, framework::ir::Node* node) {
-  if (vars_in_multi_block_map_.count(node->Name())) {
-    int var_origin_block_id = vars_in_multi_block_map_.at(node->Name()).second;
-    if (block_idx != var_origin_block_id) {
-      auto graph = graphes_[var_origin_block_id];
-      for (auto nd : graph->Nodes()) {
-        if (nd->Name() == node->Name()) {
-          return nd;
-        }
-      }
-    }
-  }
-
-  return node;
-}
-
-inline bool ConvertToMixedPrecisionPass::NodeVarHasDtype(
-    framework::ir::Node* node) {
-  if (node->IsVar() &&
-      (node->Var()->GetType() ==
-           paddle::framework::proto::VarType::SELECTED_ROWS ||
-       node->Var()->GetType() ==
-           paddle::framework::proto::VarType::LOD_TENSOR ||
-       node->Var()->GetType() ==
-           paddle::framework::proto::VarType::LOD_TENSOR_ARRAY ||
-       node->Var()->GetType() == paddle::framework::proto::VarType::STRINGS ||
-       node->Var()->GetType() == paddle::framework::proto::VarType::VOCAB)) {
-    return true;
-  }
-
-  return false;
-}
-
-// op1(fp32) -> var1, op2(fp16) -> var1
-// if and only if op1 and op2 both support fp16, we convert op1 and op2's
-// precision.
-inline bool ConvertToMixedPrecisionPass::VarIsMultiPrecisionOpsOut(
-    int block_idx, framework::ir::Node* op_node) {
-  CHECK_EQ(op_node->IsOp(), true);
-  bool ret{false};
-
-  for (auto* out : op_node->outputs) {
-    auto* real_node = GetRealNode(block_idx, out);
-    if (!real_node->Var()->Persistable() &&
-        vars_appear_multi_in_one_block_[block_idx].count(out->Name())) {
-      for (auto op_type :
-           vars_appear_multi_in_one_block_[block_idx].at(out->Name())) {
-        if (OpSupportPrecision(
-                op_type, backend_, mixed_precision_, black_list_)) {
-          ret = true;
-          VLOG(2) << out->Name()
-                  << " is multi precision op's out, so we skip convert to fp16";
-          break;
-        }
-      }
-    }
-    if (ret) break;
-  }
-  return ret;
-}
-
-void ConvertToMixedPrecisionPass::ProcessInputNode(
-    bool support_precision,
-    ir::Node* in_node,
-    ir::Node* op_node,
-    int* suffix,
-    framework::BlockDesc* block_desc,
-    framework::proto::VarType::Type to_type,
-    int block_idx) {
-  auto* real_node = GetRealNode(block_idx, in_node);
-  if (!NodeVarHasDtype(real_node)) return;
-  auto graph = graphes_[block_idx];
-  bool is_main_block = block_idx == 0;
-  auto* in_var = real_node->Var();
-  auto in_var_type = in_var->GetDataType();
-  auto prev_type = in_var_type;
-  bool is_in_multi_block = vars_in_multi_block_map_.count(in_var->Name());
-
-  if (!is_main_block && is_in_multi_block) {
-    in_var_type = vars_in_multi_block_map_.at(in_var->Name()).first;
-  }
-  if (support_precision) {
-    if (in_var->Persistable() &&
-        in_var_type == framework::proto::VarType::FP32) {
-      if (WeightsShouldNotConvert(in_node)) return;
-      in_var->SetDataType(to_type);
-      in_var_type = to_type;
-      VLOG(3) << "   in_node name " << in_var->Name() << " from " << prev_type
-              << " to " << to_type;
-    } else if (!in_var->Persistable() && IsFloatVarType(in_var_type) &&
-               in_var_type != to_type) {
-      AddCastOp(graph,
-                in_node,
-                op_node,
-                in_var_type,
-                to_type,
-                suffix,
-                block_desc,
-                &cast_map_);
-      VLOG(3) << "   in_node name " << in_var->Name() << "(" << prev_type
-              << ") to " << cast_map_[in_node]->Name() << "(" << to_type << ")";
-    }
-  } else {
-    if (!in_var->Persistable() && IsFloatVarType(in_var_type) &&
-        in_var_type != to_type) {
-      AddCastOp(graph,
-                in_node,
-                op_node,
-                in_var_type,
-                to_type,
-                suffix,
-                block_desc,
-                &cast_map_);
-      VLOG(3) << "   in_node name " << in_var->Name() << "(" << prev_type
-              << ") to " << cast_map_[in_node]->Name() << "(" << to_type << ")";
-    }
+    bool keep_io_types,
+    const std::unordered_set<std::string>& black_list)
+    : model_file_(model_file),
+      params_file_(params_file),
+      mixed_model_file_(mixed_model_file),
+      mixed_params_file_(mixed_params_file),
+      mixed_precision_(mixed_precision),
+      backend_(backend),
+      keep_io_types_(keep_io_types),
+      black_list_(black_list) {
+  if (mixed_precision_ != phi::DataType::FLOAT16 &&
+      mixed_precision_ != phi::DataType::BFLOAT16) {
+    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
+        "mixed_precision currently not supported dtype %d, we now only "
+        "support fp16 and bf16.",
+        static_cast<int>(mixed_precision_)));
   }
-}
-
-void ConvertToMixedPrecisionPass::ProcessOutputNode(
-    int block_idx,
-    ir::Node* var_node,
-    framework::proto::VarType::Type to_type) {
-  auto* real_node = GetRealNode(block_idx, var_node);
-  if (!NodeVarHasDtype(real_node)) return;
-  auto* out_var = real_node->Var();
-  auto prev_type = out_var->GetDataType();
-  if (out_var->GetDataType() == framework::proto::VarType::FP32) {
-    if (OutShouldNotConvert(var_node)) return;
-    out_var->SetDataType(to_type);
+  if (backend_ != phi::Backend::GPU) {
+    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
+        "mixed_precision currently not supported place %d, we now only "
+        "support gpu.",
+        static_cast<int>(backend_)));
   }
-  VLOG(3) << "   out_node name " << var_node->Name() << " from dtype "
-          << prev_type << " to " << out_var->GetDataType();
 }
 
-// Just process special cases.
-bool ConvertToMixedPrecisionPass::OutShouldNotConvert(ir::Node* var_node) {
-  auto op_node = var_node->inputs[0];
-  auto* op_desc = op_node->Op();
-
-  // batch_norm's input and output (variance and mean) are the same.
-  if (op_desc->Type() == "batch_norm") {
-    auto vecs = op_desc->Output("MeanOut");
-    if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-      return true;
-    }
-    vecs = op_desc->Output("VarianceOut");
-    if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-      return true;
-    }
-    vecs = op_desc->Output("SavedMean");
-    if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-      return true;
-    }
-    vecs = op_desc->Output("SavedVariance");
-    if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-      return true;
-    }
-  }
-
-  return false;
-}
+void ConvertToMixedPrecisionPass::LoadModel() {
+  framework::Executor exe{platform::CPUPlace{}};
 
-bool ConvertToMixedPrecisionPass::WeightsShouldNotConvert(ir::Node* var_node) {
-  auto op_nodes = var_node->outputs;
-  for (auto* op_node : op_nodes) {
-    auto* op_desc = op_node->Op();
-    // batch_norm op's bias, mean, scale and variance just be float32, so we can
-    // not convert the dtype.
-    if (op_desc->Type() == "batch_norm") {
-      auto vecs = op_desc->Input("Bias");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-      vecs = op_desc->Input("Mean");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-      vecs = op_desc->Input("Scale");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-      vecs = op_desc->Input("Variance");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-    } else if (op_desc->Type() == "fused_multi_transformer") {
-      auto vecs = op_desc->Input("LnScale");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-
-      vecs = op_desc->Input("LnBias");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-
-      vecs = op_desc->Input("FFNLnScale");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-
-      vecs = op_desc->Input("FFNLnBias");
-      if (std::find(vecs.begin(), vecs.end(), var_node->Name()) != vecs.end()) {
-        return true;
-      }
-    }
-  }
-
-  return false;
-}
-inline bool ConvertToMixedPrecisionPass::IsFloatVarType(
-    framework::proto::VarType::Type type) {
-  if (type == framework::proto::VarType::FP16 ||
-      type == framework::proto::VarType::FP32 ||
-      type == framework::proto::VarType::BF16)
-    return true;
-  return false;
-}
-
-void ConvertToMixedPrecisionPass::LoadAndPrepare() {
-  program_desc_ =
-      inference::Load(&executor_, &scope_, model_file_, params_file_);
+  auto program_desc = inference::Load(&exe, &scope_, model_file_, params_file_);
   main_graph_ = std::unique_ptr<framework::ir::Graph>(
-      new framework::ir::Graph(*program_desc_));
-
-  // Remove all control var
-  IrInferCleanGraphPass pass;
-  Argument arg;
-  arg.SetMainGraphNotOwned(main_graph_.get());
-  pass.Run(&arg);
-
-  vars_appear_multi_in_one_block_.resize(program_desc_->Size());
-  FindVarsInMultiBlock();
-}
-
-void ConvertToMixedPrecisionPass::FindVarsInMultiBlock() {
-  std::vector<std::set<std::string>> block_var_names_set(program_desc_->Size());
-  for (size_t i = 0; i < program_desc_->Size(); ++i) {
-    for (auto op : program_desc_->Block(i).AllOps()) {
-      auto in_names = op->InputArgumentNames();
-      block_var_names_set[i].insert(in_names.begin(), in_names.end());
-      auto out_names = op->OutputArgumentNames();
-      if (op->HasAttr("sub_block") == false) {
-        for (auto& n : out_names) {
-          if (block_var_names_set[i].count(n)) {
-            vars_appear_multi_in_one_block_[i][n].push_back(op->Type());
-          }
-        }
-      }
-      block_var_names_set[i].insert(out_names.begin(), out_names.end());
-    }
-  }
-
-  for (size_t i = 0; i < program_desc_->Size() - 1; ++i) {
-    for (size_t j = i + 1; j < program_desc_->Size(); ++j) {
-      std::set<std::string> vars_in_multi_block;
-      std::set_intersection(
-          block_var_names_set[i].begin(),
-          block_var_names_set[i].end(),
-          block_var_names_set[j].begin(),
-          block_var_names_set[j].end(),
-          std::inserter(vars_in_multi_block, vars_in_multi_block.begin()));
-
-      for (auto name : vars_in_multi_block) {
-        vars_in_multi_block_map_.emplace(
-            name, std::make_pair(framework::proto::VarType::FP32, i));
-      }
-    }
-  }
-}
-
-void ConvertToMixedPrecisionPass::ConvertAllFp64ToFp32(
-    framework::ir::Graph* graph) {
-  auto op_nodes = framework::ir::TopologySortOperations(*graph);
-  for (auto* op_node : op_nodes) {
-    if (!op_node->IsOp()) continue;
-    auto op_type = op_node->Op()->Type();
-    if (op_type == "feed" || op_type == "fetch") continue;
-
-    if (op_type == "fill_constant") {
-      if (PADDLE_GET_CONST(int, op_node->Op()->GetAttr("dtype")) ==
-          static_cast<int>(framework::proto::VarType::FP64))
-        op_node->Op()->SetAttr(
-            "dtype", static_cast<int>(framework::proto::VarType::FP32));
-    } else if (op_type == "assign_value") {
-      if (PADDLE_GET_CONST(int, op_node->Op()->GetAttr("dtype")) ==
-          static_cast<int>(framework::proto::VarType::FP64))
-        op_node->Op()->SetAttr(
-            "dtype", static_cast<int>(framework::proto::VarType::FP32));
-    } else if (op_type == "eye") {
-      if (PADDLE_GET_CONST(int, op_node->Op()->GetAttr("dtype")) ==
-          static_cast<int>(framework::proto::VarType::FP64))
-        op_node->Op()->SetAttr(
-            "dtype", static_cast<int>(framework::proto::VarType::FP32));
-    } else if (op_type == "fill_any_like") {
-      if (PADDLE_GET_CONST(int, op_node->Op()->GetAttr("dtype")) ==
-          static_cast<int>(framework::proto::VarType::FP64))
-        op_node->Op()->SetAttr(
-            "dtype", static_cast<int>(framework::proto::VarType::FP32));
-    } else if (op_type == "cast") {
-      if (PADDLE_GET_CONST(int, op_node->Op()->GetAttr("in_dtype")) ==
-          static_cast<int>(framework::proto::VarType::FP64))
-        op_node->Op()->SetAttr(
-            "in_dtype", static_cast<int>(framework::proto::VarType::FP32));
-      if (PADDLE_GET_CONST(int, op_node->Op()->GetAttr("out_dtype")) ==
-          static_cast<int>(framework::proto::VarType::FP64))
-        op_node->Op()->SetAttr(
-            "out_dtype", static_cast<int>(framework::proto::VarType::FP32));
-    }
-
-    auto inputs = op_node->inputs;
-    for (auto* in_node : inputs) {
-      auto* in_var = in_node->Var();
-      if (!in_var->Persistable() &&
-          in_var->GetDataType() == framework::proto::VarType::FP64) {
-        in_var->SetDataType(framework::proto::VarType::FP32);
-      }
-    }
-  }
+      new framework::ir::Graph(*program_desc));
+  main_graph_->SetNotOwned(framework::ir::kParamScopeAttr, &scope_);
 }
 
 void ConvertToMixedPrecisionPass::Run() {
-  LoadAndPrepare();
+  LoadModel();
 
-  for (size_t i = 0; i < main_graph_->SubGraphsSize(); ++i) {
-    auto graph = main_graph_->GetSubGraph(i);
-    graphes_.push_back(graph);
-    VLOG(2) << " --------  handle subgraph " << i << ", has "
-            << graph->Nodes().size() << " nodes --------";
+  framework::ir::AutoMixedPrecisionPass pass;
+  pass.Set("mixed_precision_mode", new int{static_cast<int>(mixed_precision_)});
+  pass.Set("mixed_black_list",
+           new std::unordered_set<std::string>{black_list_});
+  pass.Set("enable_gpu_mixed", new bool{true});
+  pass.Set("keep_io_types", new bool{keep_io_types_});
 
-    ConvertAllFp64ToFp32(graph);
-    ConvertTensorDtype(i);
-    FixCastAttr(graph);
-
-    // A trick
-    PatchForStrangeOp();
-
-    CHECK_EQ(ir::VarDescIsConsistency(*graph), true);
-  }
+  pass.Apply(main_graph_.get());
 
   SaveMixedModel();
 }
 
-void ConvertToMixedPrecisionPass::ConvertTensorDtype(int block_idx) {
-  auto graph = graphes_[block_idx];
-  framework::proto::VarType::Type to_type;
-  if (mixed_precision_ == phi::DataType::FLOAT16) {
-    to_type = framework::proto::VarType::FP16;
-  } else if (mixed_precision_ == phi::DataType::BFLOAT16) {
-    to_type = framework::proto::VarType::BF16;
-  } else {
-    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
-        "mixed_precision currently not supported dtype %d, we now only "
-        "support fp16 and bf16.",
-        static_cast<int>(mixed_precision_)));
-  }
-
-  auto op_nodes = framework::ir::TopologySortOperations(*graph);
-  auto* block_desc = op_nodes[0]->Op()->Block();
-  int num_low_precision = 0;
-  std::vector<framework::ir::Node*> output_nodes;
-
-  for (auto* op_node : op_nodes) {
-    if (!op_node->IsOp()) continue;
-    auto op_type = op_node->Op()->Type();
-    VLOG(3) << "-------------------- op_type " << op_type << ", phi_type "
-            << phi::TransToPhiKernelName(op_type);
-    // 1. set input dtype.
-    if (op_type == "feed") {
-      auto feed_var = op_node->outputs[0]->Var();
-      if (!keep_io_types_ &&
-          feed_var->GetDataType() == framework::proto::VarType::FP32) {
-        feed_var->SetDataType(to_type);
-      }
-    } else if (op_type == "fetch") {
-      auto* fetch_var = op_node->inputs[0];
-      output_nodes.push_back(fetch_var);
-      continue;
-    } else if (op_type == "cast") {
-      continue;
-    }
-
-    else if (op_node->Op()->HasAttr("sub_block")) {  // NOLINT
-      // sub_block op's output dtype should be same as input dtype, if have the
-      // same name.
-      std::unordered_map<std::string, framework::ir::Node*> in_name_to_node;
-      for (auto* in : op_node->inputs) {
-        auto* real_node = GetRealNode(block_idx, in);
-        if (NodeVarHasDtype(real_node)) {
-          in_name_to_node[in->Name()] = in;
-        }
-      }
-
-      for (auto out : op_node->outputs) {
-        auto* real_node = GetRealNode(block_idx, out);
-        if (NodeVarHasDtype(real_node)) {
-          if (in_name_to_node.count(out->Name()))
-            real_node->Var()->SetDataType(
-                in_name_to_node[out->Name()]->Var()->GetDataType());
-        }
-      }
-
-      continue;
-    }
-
-    // 2. if op support fp16/bf16 and not in blacklist.
-    //      - cast weight to fp16/bf16.
-    //      - add cast op if the input dtype is not fp16/bf16.
-    //      - set output dtype.
-    //
-    // If a var(op's out var) appears multiple times in a block, we should not
-    // convert to fp16.
-    else if (black_list_.count(op_type) == 0 &&  // NOLINT
-             !VarIsMultiPrecisionOpsOut(block_idx, op_node)) {
-      bool support_precision =
-          OpSupportPrecision(op_type, backend_, mixed_precision_, black_list_);
-
-      // if op not has float input, we will not choose the low precision kernel.
-      {
-        bool has_float_input{false};
-        for (auto in_node : op_node->inputs) {
-          auto* real_node = GetRealNode(block_idx, in_node);
-          if (real_node->Var()->GetDataType() == proto::VarType::FP16 ||
-              real_node->Var()->GetDataType() == proto::VarType::FP32 ||
-              real_node->Var()->GetDataType() == proto::VarType::FP64 ||
-              real_node->Var()->GetDataType() == proto::VarType::BF16) {
-            has_float_input = true;
-            break;
-          }
-        }
-        if (!has_float_input) {
-          support_precision = false;
-          VLOG(2) << " op doesn't has float input, just skip.";
-        }
-      }
-      VLOG(2) << " support low precision " << support_precision;
-
-      if (support_precision) {
-        VLOG(2) << " process input nodes:";
-        ++num_low_precision;
-        auto inputs = op_node->inputs;
-
-        // Just for paddle's terriable case: op's input and output has the same
-        // name.
-        std::unordered_map<std::string, std::string> names_map;
-        for (auto out_node : op_node->outputs) {
-          for (auto in_node : op_node->inputs) {
-            if (out_node->Name() == in_node->Name()) {
-              names_map[out_node->Name()] = in_node->Name();
-            }
-          }
-        }
-
-        // Process inputs.
-        for (auto* in_node : inputs) {
-          ProcessInputNode(
-              true, in_node, op_node, &suffix_, block_desc, to_type, block_idx);
-          if (names_map.count(in_node->Name()) && cast_map_.count(in_node)) {
-            names_map[in_node->Name()] = cast_map_[in_node]->Name();
-          }
-        }
-        VLOG(2) << " process output nodes:";
-        // Process outputs.
-        for (auto* out_node : op_node->outputs) {
-          ProcessOutputNode(block_idx, out_node, to_type);
-        }
-      } else {
-        auto inputs = op_node->inputs;
-        for (auto* in_node : inputs) {
-          ProcessInputNode(false,
-                           in_node,
-                           op_node,
-                           &suffix_,
-                           block_desc,
-                           framework::proto::VarType::FP32,
-                           block_idx);
-        }
-      }
-    }
-
-    // 3. check op not support fp16/bf16 or in blacklist.
-    //      - add cast op if the input dtype is not fp32.
-    else {  // NOLINT
-      VLOG(3) << "not to run fp16 op_type: " << op_type;
-      auto ins = op_node->inputs;
-      for (auto* in_node : ins) {
-        auto* in_var = in_node->Var();
-        if (in_var->GetDataType() == to_type) {
-          AddCastOp(graph,
-                    in_node,
-                    op_node,
-                    to_type,
-                    framework::proto::VarType::FP32,
-                    &suffix_,
-                    block_desc,
-                    &cast_map_);
-          VLOG(3) << "-- " << in_node->Name() << "(" << to_type << ") to "
-                  << cast_map_[in_node]->Name() << "("
-                  << framework::proto::VarType::FP32 << ")";
-        }
-      }
-    }
-  }
-
-  // 4. if output_op's dtype is not compatible to output dtype, then just
-  // insert cast.
-  for (auto* node : output_nodes) {
-    ir::Node* fetch_op{nullptr};
-    for (auto* op_node : node->outputs) {
-      if (op_node->IsOp() && op_node->Op()->Type() == "fetch") {
-        fetch_op = op_node;
-      }
-    }
-    CHECK_NOTNULL(fetch_op);
-    auto var = node->Var();
-    if (keep_io_types_ && var->GetDataType() == to_type) {
-      // fp16/bf16 -> fp32.
-      AddCastOp(graph,
-                node,
-                fetch_op,
-                to_type,
-                framework::proto::VarType::FP32,
-                &suffix_,
-                block_desc,
-                &cast_map_);
-    } else if (!keep_io_types_ &&
-               var->GetDataType() == framework::proto::VarType::FP32) {
-      // fp32 -> fp16/bf16
-      AddCastOp(graph,
-                node,
-                fetch_op,
-                framework::proto::VarType::FP32,
-                to_type,
-                &suffix_,
-                block_desc,
-                &cast_map_);
-    }
-  }
-
-  for (auto node : graph->Nodes()) {
-    auto* real_node = GetRealNode(block_idx, node);
-    if (!NodeVarHasDtype(real_node)) continue;
-
-    if (vars_in_multi_block_map_.count(real_node->Name()) &&
-        vars_in_multi_block_map_.at(real_node->Name()).second == block_idx) {
-      vars_in_multi_block_map_.at(real_node->Name()).first =
-          real_node->Var()->GetDataType();
-    }
-  }
-
-  if (num_low_precision)
-    LOG(INFO) << "---  detected " << num_low_precision
-              << " low precision ops in " << block_idx << " subgraph";
-}
-
-// We modify op's input output precision, and we need to fix cast op in_dtype
-// and out_dtype attribute.
-void ConvertToMixedPrecisionPass::FixCastAttr(framework::ir::Graph* graph) {
-  auto op_nodes = framework::ir::TopologySortOperations(*graph);
-  for (auto* op_node : op_nodes) {
-    if (!op_node->IsOp()) continue;
-    auto op_type = op_node->Op()->Type();
-    if (op_type != "cast") continue;
-    auto input = op_node->inputs[0];
-    auto output = op_node->outputs[0];
-    op_node->Op()->SetAttr("in_dtype",
-                           static_cast<int>(input->Var()->GetDataType()));
-    op_node->Op()->SetAttr("out_dtype",
-                           static_cast<int>(output->Var()->GetDataType()));
-  }
-}
-
 void ConvertToMixedPrecisionPass::SaveMixedModel() {
   framework::ProgramDesc mixed_program_desc;
   framework::ir::GraphToProgram(*main_graph_, &mixed_program_desc);
 
-  paddle::CPUPlace place;
   auto parameters = scope_.LocalVarNames();
   std::sort(parameters.begin(), parameters.end());
 
-  std::unordered_set<std::string> weights_should_be_fp32;
-  for (auto* node : main_graph_->Nodes()) {
-    if (!(node->IsVar())) continue;
-    if (NodeVarHasDtype(node)) {
-      if (node->Var()->Persistable() &&
-          node->Var()->GetDataType() ==
-              paddle::framework::proto::VarType::FP32) {
-        VLOG(2) << "weights keep to fp32: " << node->Name();
-        weights_should_be_fp32.insert(node->Name());
-      }
-    }
-  }
-
-#define CONVERT_TENSOR_DTYPE(DTYPE, dtype)                                   \
-  mixed_tensor.set_type(DTYPE);                                              \
-  auto* mixed_data = mixed_tensor.mutable_data<dtype>(platform::CPUPlace()); \
-  for (int i = 0; i < t->numel(); i++) {                                     \
-    mixed_data[i] = static_cast<dtype>(data[i]);                             \
-  }                                                                          \
-  t->clear();                                                                \
-  paddle::framework::TensorCopySync(mixed_tensor, place, t)
-
-  for (const auto& param_name : parameters) {
-    auto* var = scope_.FindLocalVar(param_name);
-    if (var->IsType<phi::DenseTensor>()) {
-      auto* t = var->GetMutable<phi::DenseTensor>();
-      if (t->dtype() != phi::DataType::FLOAT32) continue;
-      phi::DenseTensor mixed_tensor;
-      mixed_tensor.Resize(t->dims());
-      auto* data = t->mutable_data<float>(platform::CPUPlace());
-      if (mixed_precision_ == phi::DataType::FLOAT16 &&
-          !weights_should_be_fp32.count(param_name)) {
-        CONVERT_TENSOR_DTYPE(paddle::experimental::DataType::FLOAT16,
-                             phi::dtype::float16);
-      } else if (mixed_precision_ == phi::DataType::BFLOAT16 &&
-                 !weights_should_be_fp32.count(param_name)) {
-        CONVERT_TENSOR_DTYPE(paddle::experimental::DataType::BFLOAT16,
-                             phi::dtype::bfloat16);
-      }
-    }
-  }
-
-#undef CONVERT_TENSOR_DTYPE
-
   auto SerializeParams = [&]() -> std::string {
     std::ostringstream os;
     phi::CPUContext ctx;
     for (const auto& param : parameters) {
-      VLOG(3) << "Serialize param: " << param;
       PADDLE_ENFORCE_NOT_NULL(
           scope_.FindVar(param),
           platform::errors::NotFound(
               "Block should already have a '%s' variable", param));
-      auto* tensor = scope_.FindVar(param)->GetMutable<framework::LoDTensor>();
+      auto* tensor = scope_.FindVar(param)->GetMutable<phi::DenseTensor>();
       framework::SerializeToStream(os, *tensor, ctx);
     }
     return os.str();
@@ -831,96 +112,42 @@ void ConvertToMixedPrecisionPass::SaveMixedModel() {
   StrToBinary(mixed_params_file_, SerializeParams());
 }
 
-void ConvertToMixedPrecisionPass::PatchForStrangeOp() {
-  for (auto* graph : graphes_) {
-    for (auto op_node : framework::ir::TopologySortOperations(*graph)) {
-      if (op_node->Name() == "fused_multi_transformer") {
-        auto cache_kv_inputs = op_node->Op()->Input("CacheKV");
-        auto cache_kv_outputs = op_node->Op()->Output("CacheKVOut");
-        CHECK_EQ(cache_kv_inputs.size(), cache_kv_outputs.size());
-        for (size_t i = 0; i < cache_kv_inputs.size(); ++i) {
-          op_node->Op()->RenameOutput(cache_kv_outputs[i], cache_kv_inputs[i]);
-        }
-      }
-    }
-  }
+bool OpSupportPrecision(const std::string& op_type,
+                        phi::Backend backend,
+                        phi::DataType precision,
+                        const std::unordered_set<std::string>& black_list) {
+  return framework::ir::OpSupportPrecision(
+      op_type, backend, precision, black_list);
 }
-}  // namespace
 
-void AddCastOp(
+void InsertCastOp(
     framework::ir::Graph* graph,
-    framework::ir::Node* node,
-    framework::ir::Node* next_op,
+    framework::ir::Node* var_node,
+    framework::ir::Node* op_node,
     framework::proto::VarType::Type from_type,
     framework::proto::VarType::Type to_type,
-    int* suffix,
     framework::BlockDesc* block_desc,
-    std::unordered_map<framework::ir::Node*, framework::ir::Node*>* map) {
-  auto update_cast_desc = [&](framework::OpDesc& desc,
-                              const std::string& x_name,
-                              const std::string& out_name,
-                              const int in_dtype,
-                              const int out_dtype) {
-    desc.SetType("cast");
-    desc.SetInput("X", {x_name});
-    desc.SetOutput("Out", {out_name});
-    desc.SetAttr("in_dtype", in_dtype);
-    desc.SetAttr("out_dtype", out_dtype);
-    desc.SetAttr("use_mkldnn", false);
-    desc.SetAttr("with_quant_attr", false);
-    desc.Flush();
-  };
-
-  if (map->count(node) == 0) {
-    // insert cast op before node.
-    std::string cast_input_name = node->Var()->Name();
-    std::string cast_output_name =
-        node->Var()->Name() + "_cast.tmp_" + std::to_string((*suffix)++);
-    CHECK_NOTNULL(block_desc);
-    framework::OpDesc cast_op_desc(block_desc);
-    update_cast_desc(cast_op_desc,
-                     cast_input_name,
-                     cast_output_name,
-                     static_cast<int>(from_type),
-                     static_cast<int>(to_type));
-    auto* cast_op_node = graph->CreateOpNode(&cast_op_desc);
-    auto* cast_output_vardesc = block_desc->Var(cast_output_name);
-    cast_output_vardesc->SetPersistable(false);
-    cast_output_vardesc->SetDataType(to_type);
-    cast_output_vardesc->SetShape(node->Var()->GetShape());
-    auto* cast_output_node = graph->CreateVarNode(cast_output_vardesc);
-    IR_NODE_LINK_TO(cast_op_node, cast_output_node);
-    (*map)[node] = cast_output_node;
-  }
-  next_op->Op()->Rename(node->Name(), map->at(node)->Name());
-  IR_NODE_LINK_TO(node, map->at(node)->inputs[0]);
-  IR_NODE_LINK_TO(map->at(node), next_op);
-}
-
-bool OpSupportPrecision(const std::string& op_type,
-                        phi::Backend backend,
-                        phi::DataType precision,
-                        const std::unordered_set<std::string>& blacklist) {
-  auto phi_op_type = phi::TransToPhiKernelName(op_type);
-  bool support_precision = false;
-  if (blacklist.count(op_type) == 0) {
-    if (backend == phi::Backend::GPU)
-      support_precision = GpuKernelSupportPrecision(op_type, precision);
-    else
-      support_precision =
-          PhiKernelSupportPrecision(phi_op_type, backend, precision);
-  }
-  return support_precision;
-}
-
-void ConvertToMixedPrecision(const std::string& model_file,
-                             const std::string& params_file,
-                             const std::string& mixed_model_file,
-                             const std::string& mixed_params_file,
-                             phi::DataType mixed_precision,
-                             phi::Backend backend,
-                             bool keep_io_types,
-                             std::unordered_set<std::string> black_list) {
+    int* suffix,
+    std::unordered_map<framework::ir::Node*, framework::ir::Node*>* visited) {
+  framework::ir::DoInsertCastOp(graph,
+                                var_node,
+                                op_node,
+                                from_type,
+                                to_type,
+                                block_desc,
+                                suffix,
+                                visited);
+}
+
+void ConvertToMixedPrecision(
+    const std::string& model_file,
+    const std::string& params_file,
+    const std::string& mixed_model_file,
+    const std::string& mixed_params_file,
+    phi::DataType mixed_precision,
+    phi::Backend backend,
+    bool keep_io_types,
+    const std::unordered_set<std::string>& black_list) {
   ConvertToMixedPrecisionPass pass(model_file,
                                    params_file,
                                    mixed_model_file,

paddle/fluid/inference/analysis/passes/convert_to_mixed_precision.h

@@ -15,14 +15,12 @@
 #pragma once
 
 #include <string>
-#include <unordered_map>
 #include <unordered_set>
 
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/ir/graph.h"
-#include "paddle/fluid/framework/ir/graph_helper.h"
-#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/framework/scope.h"
 #include "paddle/phi/common/backend.h"
 #include "paddle/phi/common/data_type.h"
 
@@ -30,20 +28,52 @@ namespace paddle {
 namespace inference {
 namespace analysis {
 
+class ConvertToMixedPrecisionPass {
+ public:
+  explicit ConvertToMixedPrecisionPass(
+      const std::string& model_file,
+      const std::string& params_file,
+      const std::string& mixed_model_file,
+      const std::string& mixed_params_file,
+      phi::DataType mixed_precision,
+      phi::Backend backend,
+      bool keep_io_types,
+      const std::unordered_set<std::string>& black_list);
+
+  void Run();
+
+ private:
+  void LoadModel();
+  void SaveMixedModel();
+
+ private:
+  std::string model_file_;
+  std::string params_file_;
+  std::string mixed_model_file_;
+  std::string mixed_params_file_;
+  phi::DataType mixed_precision_;
+  phi::Backend backend_;
+  bool keep_io_types_;
+  std::unordered_set<std::string> black_list_;
+
+  framework::Scope scope_;
+  std::unique_ptr<framework::ir::Graph> main_graph_{nullptr};
+};
+
 bool OpSupportPrecision(const std::string& op_type,
                         phi::Backend backend,
                         phi::DataType precision,
-                        const std::unordered_set<std::string>& blacklist);
+                        const std::unordered_set<std::string>& black_list);
 
-void AddCastOp(
+void InsertCastOp(
     framework::ir::Graph* graph,
-    framework::ir::Node* node,
-    framework::ir::Node* next_op,
+    framework::ir::Node* var_node,
+    framework::ir::Node* op_node,
     framework::proto::VarType::Type from_type,
     framework::proto::VarType::Type to_type,
-    int* suffix,
     framework::BlockDesc* block_desc,
-    std::unordered_map<framework::ir::Node*, framework::ir::Node*>* map);
+    int* suffix,
+    std::unordered_map<framework::ir::Node*, framework::ir::Node*>* visited);
 
 void ConvertToMixedPrecision(const std::string& model_file,
                              const std::string& params_file,
@@ -51,8 +81,8 @@ void ConvertToMixedPrecision(const std::string& model_file,
                              const std::string& mixed_params_file,
                              phi::DataType mixed_precision,
                              phi::Backend backend,
-                             bool keep_io_types = true,
-                             std::unordered_set<std::string> black_list = {});
+                             bool keep_io_types,
+                             const std::unordered_set<std::string>& black_list);
 
 }  // namespace analysis
 }  // namespace inference

paddle/fluid/inference/analysis/passes/inference_op_replace_pass.cc

@@ -40,7 +40,7 @@ void InferenceOpReplacePass::RunImpl(Argument* argument) {
 }
 
 std::string InferenceOpReplacePass::repr() const {
-  return "inference-op-replace-pass";
+  return "inference_op_replace_pass";
 }
 
 }  // namespace analysis

paddle/fluid/inference/analysis/passes/ir_analysis_pass.cc

@@ -105,7 +105,7 @@ void IrAnalysisPass::CollectFusionStatis(Argument* argument) {
           framework::ir::kFuseStatisAttr));
 }
 
-std::string IrAnalysisPass::repr() const { return "ir-analysis-pass"; }
+std::string IrAnalysisPass::repr() const { return "ir_analysis_pass"; }
 
 }  // namespace analysis
 }  // namespace inference

paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc

@@ -64,7 +64,8 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
         "set."));
   }
 
-  auto graph = std::unique_ptr<Graph>(new Graph(argument->main_program()));
+  auto graph = std::unique_ptr<framework::ir::Graph>(
+      new framework::ir::Graph(argument->main_program()));
   argument->SetMainGraph(graph.release());
   auto *scope_ptr = argument->scope_ptr();
   PADDLE_ENFORCE_NOT_NULL(scope_ptr,
@@ -125,7 +126,7 @@ std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
   }
 }
 
-std::string IrGraphBuildPass::repr() const { return "ir-graph-build-pass"; }
+std::string IrGraphBuildPass::repr() const { return "ir_graph_build_pass"; }
 
 }  // namespace analysis
 }  // namespace inference

paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.cc

@@ -31,7 +31,7 @@ void IrGraphToProgramPass::RunImpl(Argument *argument) {
               new int(argument->memory_optim_sort_kind()));
   }
 
-  std::unique_ptr<Graph> graph(argument->main_graph_ptr());
+  std::unique_ptr<framework::ir::Graph> graph(argument->main_graph_ptr());
 
   // Direct using ProgramDesc desc(argument->main_program()) may cause
   // incomplete copies of information.

paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.h

@@ -28,7 +28,7 @@ class IrGraphToProgramPass : public AnalysisPass {
  public:
   void RunImpl(Argument *argument) override;
 
-  std::string repr() const override { return "ir-graph-to-param-pass"; }
+  std::string repr() const override { return "ir_graph_to_param_pass"; }
 };
 
 }  // namespace analysis

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc

@@ -169,7 +169,7 @@ void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
 }
 
 std::string IrParamsSyncAmongDevicesPass::repr() const {
-  return "ir-params-sync-among-devices-pass";
+  return "ir_params_sync_among_devices_pass";
 }
 
 }  // namespace analysis

paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc

@@ -295,7 +295,7 @@ void UpdateOpDescsByReuse(
   }
 }
 
-std::string MemoryOptimizePass::repr() const { return "memory optimize pass"; }
+std::string MemoryOptimizePass::repr() const { return "memory_optimize_pass"; }
 
 void MemoryOptimizePass::RunImpl(Argument* argument) {
   // Memory optimization.

paddle/fluid/inference/analysis/passes/passes.cc

@@ -18,7 +18,6 @@
 #include "paddle/fluid/inference/analysis/passes/inference_op_replace_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h"
-#include "paddle/fluid/inference/analysis/passes/ir_graph_clean_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.h"
 #include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
 #include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
@@ -34,8 +33,6 @@ PassRegistry::PassRegistry() {
                   std::unique_ptr<AnalysisPass>(new IrAnalysisPass));
   passes_.emplace("ir_graph_build_pass",
                   std::unique_ptr<AnalysisPass>(new IrGraphBuildPass));
-  passes_.emplace("ir_graph_clean_pass",
-                  std::unique_ptr<AnalysisPass>(new IrInferCleanGraphPass));
   passes_.emplace("memory_optimize_pass",
                   std::unique_ptr<AnalysisPass>(new MemoryOptimizePass));
   passes_.emplace(

paddle/fluid/inference/api/analysis_config.cc

@@ -85,15 +85,29 @@ void AnalysisConfig::SetModel(const std::string &prog_file_path,
 
   Update();
 }
+
 void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
-                                  int device_id) {
+                                  int device_id,
+                                  Precision precision_mode) {
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   use_gpu_ = true;
   memory_pool_init_size_mb_ = memory_pool_init_size_mb;
   FLAGS_initial_gpu_memory_in_mb = memory_pool_init_size_mb_;
   gpu_device_id_ = device_id;
+  mixed_precision_mode_ = precision_mode;
+  if (precision_mode == Precision::kFloat32) {
+    // default
+  } else if (precision_mode == Precision::kHalf ||
+             precision_mode == Precision::kBf16) {
+    enable_gpu_mixed_ = true;
+  } else {
+    LOG(ERROR)
+        << "The Paddle-GPU inference currently only supports "
+           "float32/float16/bfloat16 precision. Please check the parameters "
+           "you specified in EnableUseGpu or enable_use_gpu function.";
+  }
 #else
-  LOG(ERROR) << "Please compile with gpu to EnableGpu()";
+  LOG(ERROR) << "Please use PaddlePaddle with GPU version.";
   use_gpu_ = false;
 #endif
 
@@ -279,7 +293,7 @@ void AnalysisConfig::LoadIpuConfig(const std::string &config_path) {
 
     if (ipu_config_mapper_.find(key) == ipu_config_mapper_.end()) {
       PADDLE_THROW(platform::errors::InvalidArgument(
-          "invalid key {} in IPU config", key));
+          "invalid key %s in IPU config: ", key));
     }
     switch (ipu_config_mapper_.at(key)) {
       case ipu_config_code::ipu_device_num:
@@ -315,7 +329,7 @@ void AnalysisConfig::LoadIpuConfig(const std::string &config_path) {
 
       default:
         PADDLE_THROW(platform::errors::InvalidArgument(
-            "invalid key {} in IPU config", key));
+            "invalid key %s in IPU config", key));
         break;
     }
   }
@@ -372,8 +386,10 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
   CP_MEMBER(gpu_device_id_);
   CP_MEMBER(memory_pool_init_size_mb_);
 
-  // Mixed related.
+  // Mixed precision related.
   CP_MEMBER(mixed_black_list_);
+  CP_MEMBER(enable_gpu_mixed_);
+  CP_MEMBER(mixed_precision_mode_);
 
   CP_MEMBER(enable_memory_optim_);
   // TensorRT related.
@@ -740,13 +756,7 @@ void AnalysisConfig::Update() {
       ((use_custom_device() ^ pass_builder_->use_custom_device()))) {
     if (use_gpu()) {
       pass_builder_.reset(new GpuPassStrategy);
-
-      if (use_tensorrt_) {
-        // Append after the Affine_channel_conv_fuse pass.
-        pass_builder()->InsertPass(3, "tensorrt_subgraph_pass");
-      }
     } else if (use_ipu()) {
-      VLOG(1) << "IpuPassStrategy has been used for new.";
       pass_builder_.reset(new IpuPassStrategy);
     } else if (use_xpu()) {
       PADDLE_ENFORCE_EQ(
@@ -946,9 +956,6 @@ void AnalysisConfig::Update() {
         "but did not have the option -DWITH_CUSTOM_DEVICE compiled."));
 #endif
   }
-  if (ir_debug_) {
-    pass_builder()->TurnOnDebug();
-  }
 }
 
 std::string AnalysisConfig::SerializeInfoCache() {
@@ -960,6 +967,7 @@ std::string AnalysisConfig::SerializeInfoCache() {
   ss << calibration_file_path_;
 
   ss << use_gpu_;
+  ss << enable_gpu_mixed_;
   ss << use_external_stream_;
   ss << exec_stream_;
   ss << use_fc_padding_;
@@ -1167,6 +1175,7 @@ std::string AnalysisConfig::Summary() {
   os.InsertRow({"use_gpu", use_gpu_ ? "true" : "false"});
   if (use_gpu_) {
     os.InsertRow({"gpu_device_id", std::to_string(gpu_device_id_)});
+    os.InsertRow({"enable_gpu_mixed", std::to_string(enable_gpu_mixed_)});
     os.InsertRow({"memory_pool_init_size",
                   std::to_string(memory_pool_init_size_mb_) + "MB"});
     os.InsertRow(
@@ -1360,7 +1369,7 @@ bool AnalysisConfig::trt_allow_build_at_runtime() {
   return trt_allow_build_at_runtime_;
 }
 
-void AnalysisConfig::Exp_SetBlackListOpsForMixedModel(
+void AnalysisConfig::Exp_DisableMixedPrecisionOps(
     const std::unordered_set<std::string> &black_list) {
   mixed_black_list_ = black_list;
 }

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1065,7 +1065,7 @@ void AnalysisPredictor::PrepareArgument() {
   argument_.SetUseGPU(config_.use_gpu());
   argument_.SetUseFcPadding(config_.use_fc_padding());
   argument_.SetGPUDeviceId(config_.gpu_device_id());
-  argument_.SetEnableAnalysisOptim(config_.enable_ir_optim_);
+  argument_.SetEnableIrOptim(config_.enable_ir_optim_);
   argument_.SetEnableMemoryOptim(config_.enable_memory_optim());
   argument_.SetModelFromMemory(config_.model_from_memory_);
   // Analyze inference_program
@@ -1210,53 +1210,57 @@ void AnalysisPredictor::PrepareArgument() {
   }
 #endif
 
-  auto passes = config_.pass_builder()->AllPasses();
+  auto *pass_builder = config_.pass_builder();
   if (model_precision_ != phi::DataType::FLOAT32) {
     LOG(INFO) << "Model is mixed precision type with " << model_precision_
               << ", we will use a new PassStrategy. Note that only the GPU "
                  "backend is supported for now.";
-    passes.clear();
+    pass_builder->ClearPasses();
+    const auto &deleted_passes = pass_builder->GetAllDeletedPasses();
     if (config_.tensorrt_engine_enabled()) {
       for (const auto &pass : kTrtLowerPrecisionPasses) {
-        passes.push_back(pass);
+        if (deleted_passes.count(pass)) continue;
+        pass_builder->AppendPass(pass);
       }
     } else if (config_.use_gpu()) {
       for (const auto &pass : kGpuLowerPrecisionPasses) {
-        passes.push_back(pass);
+        if (deleted_passes.count(pass)) continue;
+        pass_builder->AppendPass(pass);
       }
     }
+  }
 
-    const auto &deleted_passes = config_.pass_builder()->GetAllDeletedPasses();
-    for (const auto &it : deleted_passes) {
-      auto iterator = std::find(passes.begin(), passes.end(), it);
-      if (iterator != passes.end()) {
-        passes.erase(iterator);
-      }
+  if (!config_.ir_optim()) {
+    argument_.SetEnableIrOptim(false);
+    if (config_.enable_gpu_mixed_) {
+      argument_.SetEnableIrOptim(true);
+      pass_builder->ClearPasses();
+      pass_builder->AppendPass("auto_mixed_precision_pass");
+      LOG(INFO)
+          << "This model run in Paddle-GPU mixed precision mode with no ir "
+             "optimization.";
+    } else {
+      LOG(INFO) << "ir_optim is turned off, no IR pass will be executed.";
     }
-
+  } else {
     if (config_.ir_debug_) {
-      auto it = std::begin(passes);
-      while (it != std::end(passes)) {
-        if (*it != "graph_viz_pass") {
-          it = passes.insert(it + 1, "graph_viz_pass");
-        } else {
-          ++it;
-        }
-      }
+      pass_builder->TurnOnDebug();
+    }
+    if (config_.enable_gpu_mixed_) {
+      LOG(INFO) << "This model run in Paddle-GPU mixed precision mode.";
     }
-  }
-  if (!config_.ir_optim()) {
-    passes.clear();
-    LOG(INFO) << "ir_optim is turned off, no IR pass will be executed";
   }
   argument_.SetDisableLogs(config_.glog_info_disabled());
-  argument_.SetIrAnalysisPasses(passes);
-  argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses());
+  argument_.SetIrAnalysisPasses(pass_builder->AllPasses());
+  argument_.SetAnalysisPasses(pass_builder->AnalysisPasses());
   argument_.SetScopeNotOwned(scope_.get());
 
   // mixed precison.
   argument_.SetModelPrecision(static_cast<int>(model_precision_));
   argument_.SetMixedBlackList(config_.mixed_black_list_);
+  argument_.SetEnableGPUMixed(config_.enable_gpu_mixed_);
+  argument_.SetMixedPrecisionMode(static_cast<int>(
+      paddle::ConvertPrecision(config_.mixed_precision_mode_)));
 }
 
 // NOTE All the members in AnalysisConfig should be copied to Argument.
@@ -2107,7 +2111,9 @@ std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone(void *stream) {
   }
   x->predictor_stream_ = stream;
   x->Init(scope_, inference_program_);
+#ifdef PADDLE_WITH_TENSORRT
   x->executor_->ResetTrtOps(++AnalysisPredictor::clone_num_);
+#endif
   return std::unique_ptr<PaddlePredictor>(x);
 }
 

paddle/fluid/inference/api/demo_ci/.gitignore

+data

paddle/fluid/inference/api/mkldnn_quantizer.cc

@@ -604,10 +604,8 @@ void AnalysisPredictor::MkldnnQuantizer::PrepareArgument() const {
   if (predictor_.config_.ir_debug_) builder->TurnOnDebug();
   auto passes = builder->AllPasses();
   predictor_.argument_.SetIrAnalysisPasses(passes);
-  predictor_.argument_.SetAnalysisPasses({"ir_graph_clean_pass",
-                                          "ir_analysis_pass",
-                                          "memory_optimize_pass",
-                                          "ir_graph_to_program_pass"});
+  predictor_.argument_.SetAnalysisPasses(
+      {"ir_analysis_pass", "memory_optimize_pass", "ir_graph_to_program_pass"});
   predictor_.argument_.SetQuantVarScales(scales_);
 }
 

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -247,8 +247,12 @@ struct PD_INFER_DECL AnalysisConfig {
   ///
   /// \param memory_pool_init_size_mb initial size of the GPU memory pool in MB.
   /// \param device_id device_id the GPU card to use (default is 0).
+  /// \param precision the precision used in Paddle-GPU inference.
   ///
-  void EnableUseGpu(uint64_t memory_pool_init_size_mb, int device_id = 0);
+  void EnableUseGpu(uint64_t memory_pool_init_size_mb,
+                    int device_id = 0,
+                    Precision precision_mode = Precision::kFloat32);
+
   ///
   /// \brief Turn off GPU.
   ///
@@ -967,7 +971,7 @@ struct PD_INFER_DECL AnalysisConfig {
   /// interface is in the experimental stage and may change in the future. Note
   /// that the blacklist must be the same as the model conversion blacklist.
   ///
-  void Exp_SetBlackListOpsForMixedModel(
+  void Exp_DisableMixedPrecisionOps(
       const std::unordered_set<std::string>& black_list);
 
   void SetApplyOptim(bool value) { apply_optim_ = value; }
@@ -987,13 +991,15 @@ struct PD_INFER_DECL AnalysisConfig {
   mutable std::string params_file_;
   mutable std::string calibration_file_path_;
 
-  // Mixed precision.
+  // Mixed precision related.
+  Precision mixed_precision_mode_{Precision::kFloat32};
   std::unordered_set<std::string> mixed_black_list_;
 
   // GPU related.
   bool use_gpu_{false};
   int gpu_device_id_{0};
   uint64_t memory_pool_init_size_mb_{100};  // initial size is 100MB.
+  bool enable_gpu_mixed_{false};
   bool thread_local_stream_{false};
 
   bool use_cudnn_{false};

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -171,8 +171,9 @@ const std::vector<std::string> kGpuLowerPrecisionPasses{
     "multi_devices_fused_multi_transformer_decoder_fuse_qkv_pass",
     "gpu_cpu_map_matmul_v2_to_mul_pass",
     "gpu_cpu_map_matmul_v2_to_matmul_pass",
+    "gpu_cpu_map_matmul_to_mul_pass",
     "fc_fuse_pass",
-    "fc_elementwise_layernorm_fuse_pass",
+    // "fc_elementwise_layernorm_fuse_pass",
     "embedding_eltwise_layernorm_fuse_pass",
     "runtime_context_cache_pass",
 };
@@ -227,9 +228,10 @@ GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
         "conv_elementwise_add_fuse_pass",      //
 #endif                                         //
         "transpose_flatten_concat_fuse_pass",  //
-        "constant_folding_pass",
+        "constant_folding_pass",               //
         // following pass should be located in the last, since it will
         // work on all fused ops.
+        "auto_mixed_precision_pass",  //
         "runtime_context_cache_pass"
   });